Axiomatic Reasoning for LLMs

Prompt Engineering and Requirements Engineering: Strong Structural Homology

Abstract

Requirements Engineering (RE) and the AI engineering stack of Prompt Engineering (PE), Context Engineering (CE), and Harness Engineering (HE) share a deep structural homology. Both are iterative design processes that convert vague intent into machine‑executable specifications, manage boundaries between environment and system, structure knowledge through hierarchical refinement, externalize mental models, and rely on feedback loops for validation. This report synthesizes findings across software engineering, cognitive science, and organizational psychology to delineate the common pattern and its limitations.

1. Shared Structural Principles

Five fundamental properties are common to RE and the PE–CE–HE continuum.

1.1 Transformational Pipeline

Vague stakeholder needs or ambiguous human intent are progressively converted into precise, verifiable artifacts. RE does this through elicitation → analysis → specification → verification; PE/CE/HE does it through prompting strategies → context enrichment → harness orchestration → evaluation.

1.2 Boundary Setting

Both disciplines explicitly separate the environment (given constraints, existing systems, human behavior) from the engineered artifact (software machine or AI agent’s responsibility). Jackson & Zave’s “problem frames” in RE, and the demarcation between user intent, context window content, and agent‑orchestration logic in Harness Engineering, illustrate identical boundary‑setting operations.

1.3 Hierarchical Refinement

A three‑layer abstraction is present in both.

Goal‑oriented RE (KAOS, i*) and the Spec‑Context‑Harness model map onto each other precisely.

1.4 Externalization and Shared Understanding

RE’s core challenge is externalizing distributed mental models of stakeholders into shared artifacts (vision videos, requirement documents). CE’s “Write Out” technique and HE’s Sprint Contracts externalize implicit knowledge so that multiple AI components or human‑AI teams share a common operational picture.

1.5 Design‑Scientific Iteration

Both follow a plan‑do‑evaluate cycle. RE employs validation and verification loops; HE uses Planner → Generator → Evaluator pipelines. Feedback from evaluation triggers re‑planning or re‑elicitation, making the processes structurally identical.

2. Evidence of Homology

2.1 Prompt Engineering for Requirements Engineering (PE4RE)

PE techniques (Few‑shot, Chain‑of‑Thought) are systematically applied to RE tasks such as requirement extraction, specification, and traceability, confirming that the two domains can be fused into a single methodology without architectural mismatch.

2.2 Practitioner Observations

Product managers and engineers report that crafting a good prompt maps one‑to‑one to writing a requirements document: purpose, user context, constraints, evaluation criteria, and output format are homologous elements.

2.3 Spec–Context–Harness Model

The recent three‑layer model for AI‑era development re‑organizes traditional RE into Spec Engineering (designing specifications as AI agent context) and Harness Engineering (engineering the control plane that surrounds agents). This directly mirrors the PE→CE→HE layering.

2.4 Goal Setting Theory

Organizational psychology’s finding that specific, challenging goals improve human performance aligns with the PE principle that explicit, structured prompts elicit better LLM outputs. The universality of the “specificity–performance” link supports the homology.

2.5 Boundary Objects

Both requirement specifications and structured prompts act as boundary objects that bridge disparate communities of practice (business/developers, human/AI), requiring interpretive flexibility and a common structure.

2.6 Cognitive Load Management

RE manages intrinsic/extraneous/germane cognitive load for human engineers; CE manages context‑window load for LLMs. Both are forms of limited‑capacity resource allocation, tackled with compression, selection, and decomposition.

3. Structural Correspondence Table

Dimension Requirements Engineering Prompt/Context/Harness Engineering
Primary goal Convert stakeholder needs into verifiable specs Convert intent into reliable LLM output
Target agent Human development team Single LLM call / multi‑agent system
Layering Goal → Constraint → Specification Prompt → Context → Harness
Boundary Environment (given) vs Machine (designed) User intent vs context window vs orchestration
Pipeline Elicitation → Analysis → Specification → Verification Prompt design → context curation → agent orchestration → evaluation
Externalization Shared mental models → requirement artifacts Implicit knowledge → explicit context blocks / sprint contracts
Iteration Spiral, agile feedback loops Planner‑Evaluator feedback loops
Quality dimensions Verification (internal) and validation (external) Internal code quality and external user‑need alignment
Failure modes Ambiguous reqs → wrong implementation Ambiguous prompts → irrelevant output; context degradation; self‑evaluation bias

4. Limitations of the Homology

The homology is structural, not complete identity. Key differences stem from the nature of the target agents.

These differences do not refute the homology; they refine it into a family of related design patterns parameterized by agent type.

5. Practical Implications

6. Conclusion

Requirements Engineering and the Prompt/Context/Harness Engineering stack exhibit a strong structural homology: they are both instantiations of a universal intent‑translation design pattern. This pattern encompasses transformation, boundary definition, hierarchical refinement, externalization, and iterative feedback. Differences are attributable to agent‑specific characteristics and do not undermine the underlying common architecture. Recognizing this homology enables mutual enrichment of the two disciplines and suggests a consolidated theoretical framework for designing all human‑to‑system and human‑to‑AI communication.